- immich_api: guard resp.json() with isinstance(dict) check in get_people and
fetch_all_assets so AttributeError doesn't escape on proxy/CDN non-dict responses
- executor: move actually_uploaded.append outside try/else so Frigate filename→asset_id
mapping is created via reconcile even when the tracker write fails
- cli: fall back to pre-merge people list when re-fetch after merge returns empty
(transient error) instead of silently dropping all people
- cli: treat ENABLE_FRIGATE_SCORES=false / BLUR_THRESHOLD=0 as not-set in
the unsupported-vars warning (falsy string check replaces raw truthiness)
- upload_tracker: guard set(data[flat_key]) with isinstance(list) check in
reset_person so a corrupted non-iterable legacy field doesn't crash mid-reset
- upload_tracker: guard dims[0]/dims[1] in find_by_crop_dimension with a
length check so a truncated crop_dims entry doesn't raise IndexError
- cache: wrap os.remove() in clear() with try/except OSError to handle
TOCTOU race with concurrent put() calls
- diversity: default conf_array to 0.5 (was 1.0) for faces with missing
confidence so they receive a moderate diversity boost instead of being
treated as high-confidence
- diversity: sort assets in the fast path (len <= limit) so return order is
consistent with the sorted-by-fileCreatedAt path
Correctness:
- jobs: cap auto-diversity limit for brand-new people (was never capped,
could exceed MAX_AUTO_IMAGES on first run)
- image_processing: separate None/0 guard for imageWidth/imageHeight so
missing field is explicit rather than silently aliased to img_w
- upload_tracker (_mark, update_frigate_count): copy-before-mutate so
exceptions between cache access and _save don't corrupt in-process state
- jobs: reject LIMIT=0 on no-embedding path (was silently empty run)
- jobs: add STRATEGY=skip to strategy_map so env var is honoured
- embeddings: convert to RGB before cvtColor so RGBA/grayscale thumbnails
don't raise cv2.error and silently drop from diversity selection
- config: use falsy guard for OUTPUT_DIR so blank env var falls through
to config file value
- reconcile: _ts() returns float("inf") on parse failure so unrecognised
filenames sort last instead of collapsing to 0.0 and corrupting FIFO mapping
- diversity: remove dead selected_set (never read; -np.inf sentinel already
prevents re-selection)
Lint (ruff):
- executor: sort upload_tracker import block (I001)
- executor: replace lambda is_better_than with operator.lt/gt (E731 x2)
- executor, upload_tracker: wrap long logger.warning calls (E501 x4)
- record_frigate_files_batch: copy-before-mutate so a write failure
doesn't leave cache ahead of disk (same fix as remove_frigate_files_batch)
- executor: replace tracker_ok boolean with try/else
- jobs: collapse duplicate custom_limit is not None checks into one guard
Bump version to 0.6.3.
- Drop flat list as primary storage; derive uploaded/rejected IDs from by_person
(single source of truth). Legacy flat lists in existing files still read for
backward compat. Removes dual-representation sync hazard.
- Add begin_batch/flush_batch: per-person upload loop now does 1 os.replace
instead of N (one per mark_uploaded call). Benefit on slow storage.
- reset_all_people(): RESET_PERSON=* is now O(1) disk writes instead of O(P^2).
- blur_score_from_image inlines cv2.Laplacian directly, removing assess_quality
call overhead and decoupling from the full quality pipeline.
* revert: replace SQLite tracker with JSON backend (v0.6.0)
The SQLite migration (v0.5.0) spawned 21 bug-fix releases in two days:
data-loss risk in the migration layer, schema PK conflicts on per-person
tracking, tracker isolation races under concurrent runs, and a disk-full
error that triggered duplicate Frigate uploads. The complexity cost
outweighs the benefit.
Restored the pre-SQL JSON tracker (frigate_uploaded_ids.json /
frigate_rejected_ids.json in DATA_DIR). Public API is identical — all
callers in executor.py, jobs.py, cli.py, and reconcile.py work unchanged.
Existing JSON files are read automatically; frigate_tracker.db can be
deleted once verified.
* fix: narrow corrupt-thumbnail exception to UnidentifiedImageError; restore IMMICH_URL empty-string fallback
* docs: rewrite v0.6.0 changelog, strip v0.5.x entries, fix README SQLite references
* chore: remove dead get_frigate_filename_for_asset (orphaned since FRIGATE_SCORE_THRESHOLD removal in v0.4.0)
* fix: sort imports in executor.py (ruff I001)
- diversity: cap k-medoids seed count at target so _cluster_aware_selection
never returns more images than requested (violated MAX_AUTO_IMAGES when
remaining capacity was 1-4 slots); add early return for limit=0 to
prevent k-medoids from running with a zero budget; slice return to target
as a final guard
- executor: mark_rejected() when fetch_full_image returns None so assets
that can't be fetched (both original and preview) aren't retried every run
- executor: wrap mark_uploaded() in its own try/except so a SQLite disk-full
error after a successful HTTP 200 doesn't retry the Frigate POST (duplicate
upload) — the upload succeeded; only the tracker write failed
- cli: apply skip_ids deduplication to the re-fetched people list after a
partial merge (some groups succeed, some fail) so unmerged duplicates
don't produce two jobs for the same Frigate folder
- jobs: strip whitespace from SKIP_PEOPLE/ONLY_PEOPLE elements on split
so "Alice, Bob" (space after comma) correctly matches "Bob"
- upload_tracker: replace executescript() in _migrate_schema_v2 with
individual execute() calls inside a transaction so a crash between DROP
and RENAME rolls back instead of permanently destroying tracked_assets
- frigate_api: _get_frigate_url now strips leading/trailing whitespace
before rstrip('/') so whitespace-only FRIGATE_URL is treated as unset
- executor: upload_to_frigate now uses _get_frigate_url() eliminating
double-slash upload paths when FRIGATE_URL has a trailing slash
- executor: corrupt thumbnail (resp.ok=True, Image.open fails) now calls
mark_rejected() so permanently broken assets are not retried forever
- upload_tracker: reset_person now uses _get_frigate_url() instead of
inline os.environ.get('FRIGATE_URL', '').strip()
- image_processing: _save_jpeg writes to a .tmp file and calls
os.replace() so a disk-full error never leaves a truncated JPEG
- cli: _handle_duplicate_people falls back to local deduplication when
all Immich merges fail, preventing two jobs from overwriting the same
Frigate folder
- config: _getenv_optional_float now delegates to _getenv_num() like
_getenv_optional_int, eliminating the inconsistent duplicate
- reconcile: _ts() uses rsplit('.', 1)[0] instead of .replace('.webp','')
so FIFO mapping works with any Frigate training-file extension
Correctness:
- fetch_face_data: only fall back to faces[0] when person_id is absent;
previously a missing person match injected a different person's bbox
- upload_tracker: change PK from (asset_id, status) to
(asset_id, person_name, status); old PK allowed INSERT OR REPLACE to
silently overwrite person_name when the same photo appeared in two
people's jobs, breaking quality-replacement JOINs; auto-migrates DBs
- filter_recent_assets: treat years=0 as "no age filter" instead of
falling through to Config.YEARS_FILTER via falsy `or`
- _is_module_available: return find_spec(...) is not None; find_spec
returns None (not raises) for absent top-level modules, so the
previous code always returned True
- execute_jobs error handler: use asset.get("id", "<unknown>") to avoid
a secondary KeyError propagating out of execute_jobs on malformed dicts
- upload_to_frigate: also mark_rejected on HTTP 422, not only HTTP 400
with "face" in body; other permanent errors left assets untracked and
retried forever
- reconcile_frigate_mappings: sort key lambda f: (_ts(f), f) makes order
deterministic when timestamps are equal or 0.0; set iteration order is
hash-randomised, stable sort preserves it
Reuse / cleanup:
- config.py: add _getenv_optional_int delegating to _getenv_num(name, None, int)
- jobs.py: _resolve_strategy uses _getenv_optional_int("LIMIT") instead
of inline os.environ.get + int() + warning duplicate of _getenv_num
- frigate_api.py: add _get_frigate_url() helper; eliminates 4× copy of
os.environ.get("FRIGATE_URL", "").rstrip("/")
- quality.py: extract blur_score_from_image(img, max_dim=1440) helper;
executor.py time-spread blur fallback now uses it instead of inlining
the resize+RGB+assess_quality sequence, keeping scale logic in one place
scripts/benchmark.py retained the old os.getenv inline pattern after
_getenv_bool was introduced in v0.5.16. Now uses a deferred local
import of _getenv_bool, consistent with the script's pattern of keeping
all winnow imports inside function bodies rather than at the top level.
- _getenv_num: add raw.strip() + empty-string guard so numeric vars set
to "" (common Compose pattern for "use default") return the default
silently instead of warning "not a valid int/float"
- _getenv_bool: same guard so True-defaulted flags set to "" return
the configured default instead of silently returning False
- embeddings.py: replace inline FORCE_CPU bool parse with _getenv_bool
- cache.py: fix np.save extension bug from v0.5.13 — tmp path used
final+".tmp" (abc.npy.tmp) but np.save auto-appends .npy to paths not
ending in .npy, writing to abc.npy.tmp.npy instead; os.replace then
raised FileNotFoundError silently, making every cache write a no-op
and leaking *.npy.tmp.npy files. Fixed by inserting .tmp before .npy:
tmp = final[:-4] + ".tmp.npy"
- config.py: remove str(default) round-trip in _getenv_int/_getenv_float
— use raw = os.getenv(name); return default if raw is None else int(raw)
so a future float default can't cause a spurious "not a valid integer"
warning and return the wrong type
- executor.py: consolidate 4 progress.remove_task calls into one
try/finally around the per-job body; continue inside try/finally
executes the finally before the next iteration, making the invariant
structurally enforced rather than relying on discipline across 4 sites
YEARS_FILTER, MIN_FACE_WIDTH, MIN_FACE_COUNT, MAX_AUTO_IMAGES,
BLUR_THRESHOLD, MIN_CONFIDENCE, and FACE_MARGIN used bare int()/float()
calls with no error handler. A typo (trailing space, non-numeric value)
raised ValueError inside __getattr__, producing a cryptic traceback on
the first config access rather than at the validate() step. Values are
now parsed by _getenv_int/_getenv_float helpers that warn and fall back
to the documented default, matching the existing FRIGATE_SCORE_CEILING
pattern.
- executor.py: wrap shutil.rmtree/os.makedirs in try/except OSError so a
permission failure logs and skips the job rather than aborting the run
- cache.py: write embeddings to a .tmp file and atomically rename into place
via os.replace so a process kill can't leave a corrupted .npy cache slot
- immich_api.py: guard fileCreatedAt with isinstance(str) check before calling
.replace() so a non-string timestamp doesn't raise AttributeError and kill
the entire filter_recent_assets pass
- upload_tracker.py: raise SQLite busy timeout from 5 s to 30 s to handle
concurrent cron+manual run overlap without dropping upload-tracking records
progress.add_task() fires unconditionally at the top of the job loop;
both continue paths (ValueError from _safe_person_dir and the symlink
TOCTOU guard) skipped remove_task(), leaving orphaned 0% rows in the
terminal for the rest of the run.
The v0.5.10 compound guard 'isdir and not islink' silently skipped the
rmtree when person_dir was a symlink-to-directory, then let makedirs
follow the symlink — allowing crop writes outside output_dir with no
diagnostic. Replace with an explicit islink pre-check that logs an error
and continues, matching the ValueError path from _safe_person_dir.
- reconcile.py: re-escalate the < target branch from INFO to WARNING and
add 'permanently unmapped' label. Both post-loop branches produce identical
permanent mapping loss; v0.5.9 incorrectly treated the timeout case as
recoverable.
- executor.py: guard shutil.rmtree with 'not os.path.islink(person_dir)'
so a race-replaced symlink-to-directory is skipped rather than raising
an unhandled OSError that aborts all remaining jobs. Correct comment:
rmtree raises OSError, not NotADirectoryError.
- reconcile.py: swap log levels — external-upload path (permanent mapping
loss) escalated to WARNING; timeout path (transient, retries next cycle)
downgraded to INFO. Also extend the warning message to note the files are
permanently unmapped.
- immich_api.py: extend fetch_all_assets docstring to document that
all-garbage page termination (in addition to network errors) makes
total_raw a lower bound.
- executor.py: add comment above shutil.rmtree noting that POSIX rmtree
raises NotADirectoryError on a top-level symlink, documenting why the
removed islink guard is safe to omit.
- upload_tracker.py: replace setdefault with explicit guard in _entry() —
setdefault evaluates its default-dict argument before checking key
presence, allocating and discarding a dict on every already-present call.
- executor.py: fix _safe_person_dir docstring — realpath+startswith is the
load-bearing traversal guard; islink is a supplementary early-exit for the
symlink sub-case only. The previous comment "checking after realpath would be
too late" implied islink was the primary guard, which is backwards.
- immich_api.py: move total_raw accumulation to after the dead-end-page break
so all-garbage pages don't inflate the count and produce misleading
"N total, 0 recent" output. Mixed pages (some valid, some non-dict) still
count page_count so transient schema issues don't shrink MIN_FACE_COUNT below
threshold. Add warning when a RequestException interrupts pagination mid-way
so operators know total_raw is a lower bound.
- config.py: eliminate residual TOCTOU — change `if config_file.exists():` to
`if _data_cfg_exists or config_file.exists():` so _data_cfg is never
stat'd twice (the v0.5.7 fix cached the first check but not the second).
executor.py:
- Move islink check into _safe_person_dir on the raw path, before realpath
resolves it; the previous check at the rmtree site was unreachable dead code
because realpath already followed any symlink
immich_api.py / jobs.py:
- fetch_all_assets now returns (assets, total_raw) where total_raw is the
item count seen before non-dict filtering; callers use it for MIN_FACE_COUNT
guard and display so transient non-dict API items can't incorrectly skip people
- Add WARNING when pagination stops because a page had items but all were non-dict
config.py:
- Cache _data_cfg.exists() in _data_cfg_exists so the dual-config warning
and config_file selection always read from the same stat() result; previously
two calls created a TOCTOU window where log and code could disagree
Bump version to 0.5.7
immich_api.py:
- Move empty-page break after non-dict filtering — a page of all-null
items no longer loops to MAX_PAGES without terminating
- Single-pass partition replaces two inverse isinstance scans per page
- Upgrade non-dict item log from DEBUG to WARNING (silent asset loss)
reconcile.py:
- Check Frigate before the first sleep so fast responses return
immediately rather than always paying a 1 s delay
- Compute set difference once per poll iteration instead of twice
Bump version to 0.5.6
* fix: quality replacement slot floor uses deleted file's score not failed candidate's
When a blur-score replacement deletes a low-quality Frigate file but the
subsequent upload fails, min_quality_score_for_slot was set to candidate_score
(the good file that failed to upload). This filtered out any subsequent
candidate that didn't beat the failed upload, even if it was better than
the file we just deleted — leaving the freed slot unfilled unnecessarily.
The comment on the guard already documented the correct intent: 'require
the next candidate to beat the deleted file's score'. Fix: use target_score
(the deleted file's blur score) as the floor instead of candidate_score.
* chore: bump version to 0.5.4
* fix: audit hardening — input validation, error handling, and robustness
- immich_api: guard person["id"] with .get() + early return on missing field
- immich_api: include page number in pagination exception log
- immich_api: validate faces response is a list before indexing
- executor: wrap Image.open() in try/except for non-image HTTP responses
- executor: strip leading 'v' from Frigate version before parsing (v0.16.0 was misread)
- config: wrap FRIGATE_SCORE_CEILING float() parse in try/except with warning
- config: warn when both DATA_DIR and legacy CWD config files exist simultaneously
- scheduler: wrap PID file write in try/except so /tmp failures don't crash startup
- scheduler: clamp sleep to 60s max to bound recovery time after NTP clock jumps
- frigate_api: log unexpected non-list type in get_frigate_person_files at DEBUG
* fix: LIMIT env var crash and symlink guard on person output dir
- jobs: wrap int(LIMIT) parse in try/except — bad value (e.g. "30.5", "all")
now logs a warning and falls back to the default instead of crashing
- executor: check for symlink before shutil.rmtree on person_dir — prevents
following a symlink out of OUTPUT_DIR on a shared volume
* chore: bump version to 0.5.3
* refactor: rename CACHE_DIR to DATA_DIR, default path .if_cache → data
CACHE_DIR held both the embedding cache and the SQLite tracker DB, making
the name misleading. DATA_DIR is more accurate.
- Config reads DATA_DIR first; falls back to CACHE_DIR with a deprecation
warning so existing setups don't break on upgrade
- Default local path: data (was .if_cache)
- Docker default path: /app/data (was /app/.if_cache)
- Internal references (embeddings.py, upload_tracker.py) updated to DATA_DIR
- compose.yml, .env.example, README, wiki, and changelog updated
- Version bumped to 0.5.1
* chore: update lockfile
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* refactor: rename CACHE_DIR to DATA_DIR, default path .if_cache → data
CACHE_DIR held both the embedding cache and the SQLite tracker DB, making
the name misleading. DATA_DIR is more accurate.
- Config reads DATA_DIR first; falls back to CACHE_DIR with a deprecation
warning so existing setups don't break on upgrade
- Default local path: data (was .if_cache)
- Docker default path: /app/data (was /app/.if_cache)
- Internal references (embeddings.py, upload_tracker.py) updated to DATA_DIR
- compose.yml, .env.example, README, wiki, and changelog updated
- Version bumped to 0.5.1
* chore: update lockfile
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
- release.yml: replace per-variant lockfile generation loop with a
single 'uv lock'; add idempotent release creation (skip if tag
already has a release so re-triggered runs don't 422)
- docker-publish.yml: remove stale uv-cpu/rocm/intel.lock entries
from paths-ignore (those files no longer exist)
- README: CACHE_DIR description now names winnow_tracker.db; tracker
description mentions SQLite
Two fixes found during post-refactor audit:
1. upload_tracker: remove COUNT(*) guard from _maybe_migrate. The guard
blocked re-migration when a previous run successfully committed both
JSON files but a PermissionError on the second rename() left it on
disk. On the next startup COUNT > 0 → early return → rejected IDs
permanently unimported. INSERT OR IGNORE is idempotent so re-running
migration is always safe; guard not needed.
Also wrap each rename() in its own try/except so a failure on one
file is logged and does not propagate uncaught.
2. reconcile: break early when new_count > target is detected in the
poll loop. Previously the loop ran all four delay intervals (1+2+4+8s)
before the post-loop > target branch fired, wasting up to 15 seconds
when a concurrent external upload was visible on the first poll.
- upload_tracker: partial migration now rolls back atomically on failure;
JSON renamed only after successful commit so failed runs retry cleanly
- upload_tracker: allowlist score_col in _pick_mapped_file to close
latent SQL injection surface
- config: move load_dotenv() from module import into _load() so no I/O
at import time and reset() fully resets env loading
- config: use is None checks for IMMICH_URL/OUTPUT_DIR config-file
fallback so explicitly empty env vars are not overridden by the file
- executor: skip reconcile when Frigate API is unreachable at upload
start — tracker baseline is incomplete and would mis-trigger the
external-upload guard, permanently losing file mappings
- reconcile: change polling break condition from >= to == target so
transient overshoots don't prematurely exit the loop and trigger
the external-upload guard
- jobs: apply capacity cap as the selection limit rather than truncating
post-selection by position, so the diversity algorithm works within
the right budget from the start
- Dockerfile: explicit gpu branch + exit 1 on unknown VARIANT instead
of silent fallback
_build_job no longer calls filter_already_uploaded internally; callers pass
pre-filtered assets so there's no double DB hit and the interactive path
restores the original prompt order (retry_rejected asked before strategy,
so post-dedup count informs the choice). Skip-count rprint restored in
auto_configure. Late 'import time' inside reconcile_frigate_mappings moved
to module level.
- rocm/intel/gpu extras are x86_64-only; aarch64 wheels don't exist so uv
failed to resolve them when required-environments includes aarch64
- test.yml: switch from --all-extras (broken by conflicts + missing wheels)
to --extra cpu which is cross-platform and sufficient for unit tests
- update-lockfile.yml: drop old file-swap loop; single pyproject means a
single uv lock run and a single uv.lock to commit
- Config: remove _ConfigAccessor and ConfigManager; use __getattr__ for lazy
loading on single _Config class; re-register self as _instance in __getattr__
so reset() always clears the correct object (item 1)
- upload_tracker: replace hand-rolled JSON store with sqlite3; auto-migrates
existing JSON on first run; remove dead record_frigate_file function;
connection re-opens when CACHE_DIR changes for test isolation (items 2, 8)
- diversity: move ThreadPoolExecutor import to module level; inject optional
fetch_fn parameter for testability (items 3, 6)
- pyproject: consolidate 4 variant files into extras (gpu/rocm/intel/cpu);
update Dockerfile to use --extra flag; delete variant pyproject/lock files;
uv.lock needs regen with `uv lock` after this change (item 4)
- jobs: extract _build_job helper to separate business logic from terminal I/O;
auto_configure delegates dedup/selection to _build_job (item 5)
- logging: convert f-string log calls to % interpolation throughout all winnow/
modules (item 7)
- reconcile: new module with reconcile_frigate_mappings and
enrich_asset_with_face_data extracted from executor.py (item 9)
- scheduler: print next scheduled run time after startup and after each run;
fix f-string logger.error call (item 10)
22.04 non-GPU amd64 bases were never updated when arm64 moved to 24.04.
Python 3.13 is still pulled from deadsnakes PPA (26.04 ships 3.14 natively).
intel stays on 22.04: the Intel GPU repo URL is pinned to the "jammy"
codename and cannot be bumped until Intel publishes 26.04 packages.
cache.py — model fingerprint auto-invalidation:
Replace hardcoded "buffalo_l_v1" version string with a fingerprint
derived from buffalo_l .onnx file sizes and mtimes. EmbeddingCache now
computes this at init time; stale embeddings from replaced or updated
model files are automatically invalidated. Falls back to the static
string before the model is downloaded.
Note: existing caches built against the old key will miss on the first
run after upgrade and recompute cleanly.
frigate_api.py — Frigate version check:
Add get_frigate_version() (GET /api/version). Called at the start of
upload_to_frigate(); warns if below v0.16 where the face training API
endpoints don't exist.
immich_api.py + cli.py — Immich version check:
Add get_immich_version() (GET /api/server/version). Called at startup
before get_people(); warns if below v1.106 where the face data and
merge APIs winnow depends on aren't guaranteed present.
Remaining TODO(frigate-api) annotations are left in place — they require
Frigate to expose per-file embeddings or a rebuild-complete signal before
they can be addressed.
Adds inline LIMITATION / TODO(frigate-api) comments at each specific
code site rather than a separate doc that would drift from the code.
frigate_api.py — recognize_face:
Mean-embedding limitation: score reflects the arithmetic mean of all
training embeddings. A bimodal set (frontals + profiles) has a mean
between clusters, making both ends look more novel than they are.
Fixable if Frigate exposes per-file embeddings for nearest-neighbour
comparison.
frigate_api.py — get_all_frigate_person_files:
"train" key exclusion is a hardcoded string. If Frigate adds other
special top-level keys in /api/faces they'll be silently treated as
person names. Needs a typed schema when Frigate documents the contract.
executor.py — recognize_face call site:
Async rebuild: each deletion triggers a background model rebuild in
Frigate. Subsequent recognize calls in the same run return None
(rebuild in progress), degrading quality replacement for later
candidates. Fixable with a rebuild-complete signal from Frigate.
executor.py — effective_count / manual file handling:
Manually-added files are invisible to diversity decisions. Winnow
observes their effect only indirectly via the Frigate score, not by
measuring their embedding distribution. Per-file embeddings from
Frigate would allow direct diversity measurement against the full set.
executor.py — Frigate version assumption:
All face training endpoints are v0.16+. No version check at startup;
failures on older versions are opaque 404s.
cache.py — MODEL_VERSIONS:
Version string is a hardcoded constant. Manual model file replacement
(custom weights, InsightFace update) won't invalidate cached embeddings.
Needs file-checksum-derived versioning or a CLEAR_EMBEDDING_CACHE flag.
diversity.py — thumbnail-resolution embeddings:
Diversity selection runs InsightFace on preview thumbnails; the actual
training crop comes from full-resolution originals. Negligible in
practice but degrades if Immich preview quality is low.
Tests cover the core ML pipeline algorithms in diversity.py — previously
untested. No network or model dependencies; all pure-function or
numpy-only paths:
- Face bbox and confidence extraction from Immich metadata, including
person_id filtering and missing-data edge cases
- Face crop scaling: verifies bbox coordinates are correctly scaled when
the thumbnail dimensions differ from the metadata image dimensions
- Near-duplicate dedup: removal below cosine threshold, quality-score
preference between duplicates, zero-quality-score treated as zero not
missing (falsy bug guard)
- K-Medoids: correct medoid count, distinctness, valid index range, and
full-N edge case
- Adaptive threshold: positive output, floor at 0.05 for identical
embeddings, single-point, scales with embedding spread
- Time-spread fallback: exact count, all-under-limit passthrough,
auto→30 default, first/last inclusion
- Cluster-aware selection: exact limit, subset invariant, auto-stop on
tight cluster, hard-example confidence weighting accepted
Remove torch/torchvision/transformers/ultralytics from all three variant
pyproject files (rocm/cpu/intel) — no longer needed since SigLIP and YOLO
were removed in v0.4.10. Update version to 0.4.10 and description.
NOTE: uv-rocm.lock, uv-cpu.lock, uv-intel.lock are now stale and must be
regenerated in their respective platform environments before the next Docker
build of those variants.
Also fix fetch_face_data() docstring to not mention embedding (removed field).
- immich_api.py: remove FaceData.embedding field and numpy import — Immich face
embeddings were never consumed after being fetched; bbox/confidence is all that's used
- embeddings.py: remove immich_embedding param from get_embedding() — never passed by
any caller; simplify docstring accordingly
- image_processing.py: wrap insightface_app.get() in warnings filter to suppress the
scikit-image FutureWarning about estimate being deprecated (already suppressed in
embeddings.py for the diversity path, was leaking from the crop-alignment path)
- README.md: remove two stale "object mode" / "object classification" fragments
After removing the object pipeline, several dead 'mode' artifacts remained:
- executor.py: unpack `config` from job even though it was no longer read
- jobs.py: set `"mode": "face"` in both configure paths (key never consumed)
- compose.yml: TRAINING_MODE=face env, OBJECT_CLASS comment, HF_HOME, stale MAX_AUTO_IMAGES default note
- Dockerfile: "and object classification" label, /models/huggingface mkdir, HF_HOME ENV
- cache.py: drop siglip MODEL_VERSIONS entry
- log_config.py: drop ultralytics/transformers/torch from noise silencers
- scheduler.py: drop HF_HOME and HuggingFace model check
- scripts/benchmark.py: drop bench_siglip and make_random_image
winnow is a face recognition training tool. Object mode required manual
file placement with no Frigate API, pulled in torch/torchvision/transformers/
ultralytics (~2 GB), and was architecturally misaligned with the project goal.
Removed:
- process_object_mode (YOLO inference), get_yolo_model
- SigLIP model stack (get_siglip_model, get_object_embedding, batch variant)
- entity_type branching throughout diversity, embeddings, jobs, executor
- TRAINING_MODE and OBJECT_CLASS env vars
- torch, torchvision, transformers, ultralytics dependencies
- pytorch index entries from pyproject.toml
- Object mode from README (Modes section, env var table, How It Works)
- How It Works step 9: "upload freely" → note novelty gate may skip below-cap candidates
- Persistence note: "Frigate rejections" → "rejected assets" (covers confidence skips too)
- RETRY_REJECTED description: explicitly covers all rejection types, not just Frigate
- Image Quality section: split into user-adjustable controls and calibrated image
processing defaults with a support disclaimer to deter blind tuning
flan
